Web Search Sop

Web Search SOP

Skill for the deep-research search sub-agent: how to issue web_search queries that maximise official-source hit-rate while keeping the call count to one.

When to apply

The root DeepAgent dispatches a natural-language question. Treat it as a single search task — no clarification, no follow-up dispatch.

Step 1 — Detect query language and locale hints

• If the question contains CJK characters, set language: "zh".
• If the question contains only Latin script, set language: "en".
• Mixed: prefer the script of the proper nouns; default to "zh" for product names that originate in Chinese (豆包, 通义, 文心, etc.).

Step 2 — Choose time_range

• Question mentions "最近"/"latest"/"2026"/this year → time_range: "month".
• Question references a known release window (e.g. "DeepSeek V3.1 发布") → time_range: "year".
• General factual lookup → time_range: "all".

Step 3 — Build the query string

• Pass through verbatim first. Tavily already understands natural language.
• Keep one vendor brand name if present; strip generic prefixes like "请问"/"can you tell me".
• Quote multi-word phrases only when the question demands an exact phrase ("agent runtime" not agent runtime).

Step 4 — Pick top_k

• Comparison / list questions → top_k: 8.
• Single-fact lookup (a version number, a release date) → top_k: 5.
• Never exceed 10 — Tavily charges per result and the downstream verifier deduplicates anyway.

Step 5 — Retry rule

Call exactly once. Retry only if results is empty. Retry strategy:

1. Drop the vendor brand name from the query.
2. Switch language to "any".
3. Do not change top_k or time_range.

Never retry a non-empty result. Never make two calls in a single turn.

Output contract reminder

Return the tool's JSON output as-is. Do not rewrite snippets, do not invent fields, do not synthesise summaries. The root agent owns verification and synthesis.